When a tool-driven voice adapter starts allowing text-only backend replies, consume caller

When a tool-driven voice adapter starts allowing text-only backend replies, consume caller transcript fragments when those replies complete as well as when tools execute. Otherwise a clarification or reminder stays queued and is concatenated into the next caller turn, corrupting latest-turn grounding and retry accounting. Bind work to a known delegation snapshot, consume its input once, retain genuinely later fragments, and reject tool execution when there is no fresh caller input or no known delegation. A model prompt alone is not that execution boundary.

Applicability and limits: This applies to adapters that buffer input transcript deltas and previously used tool execution as the only turn-completion event. It is separate from deciding semantic utterance boundaries and does not justify using unrelated historical digits or treating backend completion as proof of finished audible speech.